Automobile risk control time sequence prediction method based on large model technology

Abstract

The invention is suitable for the technical field of automobile risk control prediction, and provides an automobile risk control time sequence prediction method based on a large model technology, and the method comprises the steps: carrying out the fusion and serialization processing of multi-source heterogeneous data, such as credit video data, vehicle GPS data, regular return visit audio data, and repayment behavior data; a comprehensive time sequence data basis is constructed; according to the method, a comprehensive, accurate and robust time sequence prediction framework is constructed, efficient, accurate and timely prediction of customer risk behaviors is realized, the data redundancy is remarkably reduced, the data processing and model training efficiency is improved, the universality and adaptability of the model in different scenes are enhanced, and the method is suitable for being popularized and applied. Meanwhile, dependence on historical data is reduced, potential risk modes can be mined in time, risk early warning decision support with higher timeliness and accuracy is provided for financial institutions, and therefore the practicability and economic benefits of the model are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automobile risk control prediction, and particularly relates to an automobile risk control time series prediction method based on large model technology. BACKGROUND

[0002] With the booming development of the automobile finance market, the demand for financial services such as automobile loans and leases has grown rapidly. However, behind this prosperity, the risks faced by the automobile finance industry have become increasingly prominent, including credit risk, fraud risk, and vehicle value fluctuation risk. These risks are not only complex and variable, but also pose a serious challenge to the stable operation of financial institutions.

[0003] In traditional automobile finance risk control practices, risk prevention and control measures mainly rely on the experience of business experts to develop risk control rules or use historical blacklists for risk identification. Although these methods can play a role to some extent and have high interpretability, they have significant limitations. The experience of business experts is often limited by the time range and is difficult to fully cover all potential risk patterns. Blacklists have timeliness and privacy issues and cannot reflect the latest risk dynamics in a timely manner. More importantly, these methods do not take into account the changes in financial risks over time, and their occurrence patterns also change, making it difficult to achieve timely tuning and adaptation, and making them less effective in dealing with new financial risks.

[0004] With the rapid development of artificial intelligence technology, methods based on machine learning and deep learning have been gradually introduced into the field of risk control in order to improve the timeliness and accuracy of risk identification. These methods can effectively identify potential financial risks by mining potential patterns in data. However, they also face some challenges. First, these methods usually require a large amount of labeled data for training, and the determination of labels often requires a certain customer performance period, which leads to a certain lag in these methods when faced with changes in financial risk patterns over time, making them unable to fully adapt to new risk environments. Second, traditional methods often have difficulty effectively fusing and extracting key information from multi-source heterogeneous data, limiting the accuracy and robustness of prediction models. SUMMARY

[0005] The present application provides an automobile risk control time series prediction method based on large model technology, which aims to solve the problem that existing risk control methods rely on expert experience and blacklists, have timeliness and privacy issues, and cannot adapt to dynamic changes in risk patterns. Methods based on machine learning and deep learning can improve risk identification accuracy, but rely on a large amount of labeled data, have a lag, and have difficulty effectively fusing and extracting key information from multi-source heterogeneous data, limiting the accuracy and robustness of prediction models.

[0006] The application is implemented as follows: a vehicle risk control time series prediction method based on large model technology, comprising S1: fusing and serializing multiple source heterogeneous data to obtain time series data; the multiple source heterogeneous data comprises credit investigation video data, vehicle GPS data, regular return audio data and repayment behavior data;

[0007] S2: constructing a time series prediction model based on an attention mechanism, receiving past P time point input context of the time series data, and outputting overdue prediction results of future F time points;

[0008] S3: optimizing the sampling strategy of the time series prediction model through a reinforcement learning method to improve the prediction performance of the model.

[0009] Preferably, step S1 comprises the following sub-steps:

[0010] S11: adaptively sampling and fixedly sampling the credit investigation video data to obtain candidate video frames;

[0011] S12: extracting stay points and performing multi-modal conversion on the vehicle GPS data to obtain GPS text data and GPS image data;

[0012] S13: extracting features from the regular return audio data to obtain audio feature data;

[0013] S14: standardizing text conversion on the repayment behavior data to generate repayment record text;

[0014] S15: aligning modalities and time of image, text and audio data to form a unified time series input.

[0015] Preferably, step S11 comprises:

[0016] The overlap rate of adjacent video frames is calculated, and given two video frames F i and F j , the calculation formula of the overlap rate is as follows:

[0017] Preferably, step S12 comprises:

[0018] Stay point extraction: based on a time threshold and a distance threshold, all points meeting the given threshold condition in the n-day GPS points are selected as the candidate stay point set of the user;

[0019] Stay point fusion: the mean fusion is performed on the continuous GPS points of the same stay point area, that is, the mean fusion is performed on the stay points within the same distance threshold range as the new stay point starting from the first stay point based on the time series order, and finally the stay point set of the user in the n days is obtained.

[0020] Multi-modal conversion: concatenate the n-day stay point set into GPS text data, and draw it into a GPS heat map image to obtain GPS image data.

[0021] Preferably, step S13 comprises:

[0022] The audio data is preprocessed by noise reduction and frame division, and the mel frequency cepstral coefficient and short-time energy spectrum are calculated to obtain the audio feature tensor.

[0023] Preferably, step S14 comprises:

[0024] The repayment record is converted into structured text in the format "timestamp: repayment amount→actual repayment amount", and the repayment behavior text data is generated.

[0025] Preferably, step S15 comprises:

[0026] Time alignment: based on the second-level timestamp, the video frame, GPS point, audio segment and repayment record are mapped to a unified time axis, and the missing values are completed by the context prediction of the large model;

[0027] Block serialization: the multi-modal data is divided into blocks according to a fixed time window, each block is spliced into a multi-modal sequence in time sequence, and position encoding is added to retain time sequence information.

[0028] Preferably, the time series prediction model of step S2 is based on a large model architecture, specifically comprising:

[0029] A time series prediction model is constructed, which takes the data of the past P time points in the time series as input context information to predict the potential delinquency status of the customer at the future F time points:

[0030]

[0031] The prediction accuracy is evaluated by the metric corresponding to the current delinquency day index of the repayment behavior data, which quantifies the closeness between the predicted value and the true value, and the mean absolute error is used as this metric:

[0032]

[0033] Step 1. In the time series preprocessing link, to generate the input token required by the transformation layer, first divide the input time series data into continuous and non-overlapping segments, and after each segment is processed by a residual block, it is converted into a vector adapted to the model dimension. At the same time, a binary format padding mask is introduced, where 1 represents the corresponding input value is masked, and 0 represents it is not masked. After the above processing, the input data is divided into segments of length n, and the i-th segment is denoted as Corresponding mask segmentation is recorded as The i-th input token is represented as:

[0034]

[0035] Where PE i represents the i-th position encoding, represents element-wise multiplication, and there are a total of input tokens;

[0036] Step 2. Design a stacked transformer layer as the core structure of the model, using a standard causal multi-head self-attention mechanism to ensure that each output token can only pay attention to the input tokens that precede it in the sequence when performing attention calculations. The formula is as follows:

[0037]

[0038] Where ST represents the stacked transformer layer, t i represents the i-th position token, and the i-th represents the minimum mask for the i-th position.

[0039] Step 3. Design the output layer to map the output token to the predicted value. The model is trained in a decoder mode to ensure that each output token predicts the time series part that follows the last input segment corresponding to itself. Set the output segment length to h, and use another residual block to map the output token to the predicted value. The formula is as follows:

[0040]

[0041] As above, all data encoding in y 1:ni is completed, and it is used to predict the next h time points of the sequence y ni+1:ni+h .

[0042] This model aims to predict the number of overdue days in the repayment behavior data. During training, the mean squared error is used as the loss function to measure the difference between the predicted value and the true value and optimize the model:

[0043]

[0044] W is the number of samples in a batch, is the predicted value of the next segment of the time series by the model, y ni+1:ni+h is the corresponding actual observed value.

[0045] Preferably, step S3 includes the following sub-steps:

[0046] S31: define a set of time series sampling strategies, including weighted random sampling: let w(t) = mu (T-t) is the weight of the data point at time t, where w(t) = mu (T-t) is close to but less than 1, T is the selected prediction time point, and the sampling probability is defined as

[0047] S32: build a reinforcement learning framework: the state space S is the prediction accuracy of the model within the time window; the action space A is the set of selectable sampling strategies;

[0048] The reward function is:

[0049] where A t represents the accuracy of the model in the time window [t, t+Delta t], A threshold represents a preset accuracy threshold, and alpha represents a hyperparameter that controls the slope of the reward function.

[0050] The policy function is:

[0051] where a t represents the action selected at time step t, and s t represents the state at time step t, i.e. the accuracy of the sample prediction in the time window [t, t+Delta t];

[0052] S33: update the policy through a deep Q network, and the algorithm formula is as follows:

[0053]

[0054] S34: dynamically select the sampling strategy according to the optimized policy, retrain the time series prediction model until the model accuracy converges.

[0055] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0056] Firstly, the present application builds a comprehensive and accurate time series prediction framework through multi-source heterogeneous data fusion, fine time series preprocessing, attention mechanism modeling, and reinforcement learning dynamic optimization of sampling strategies, significantly improving the accuracy, timeliness and robustness of the prediction. Specifically, multi-source data fusion captures the multi-dimensional features of customer behavior, preprocessing and attention mechanism accurately extract the complex relationships in the time series, and reinforcement learning dynamically adjusts the sampling strategy according to the real-time prediction effect, ensuring that the model can efficiently utilize data resources under different time windows, ultimately achieving high precision and high timeliness of the prediction result, and providing reliable and intelligent decision support for the automobile risk control field.

[0057] Secondly, the application converts multi-source heterogeneous data into unified time sequence input through an efficient data processing method, combines a reinforcement learning framework with a deep Q network to dynamically optimize a sampling strategy, and realizes efficient, accurate and timely prediction of customer risk behaviors. This method not only reduces data redundancy, improves data processing efficiency, but also significantly improves model training efficiency and prediction accuracy. At the same time, its dynamic optimization and adaptive ability ensures the universality and adaptability of the model in different scenarios. More importantly, this method reduces the dependence on historical data, can timely mine risk patterns and predict customer risk behaviors, provides financial institutions with more timely and accurate risk warning decision support, and significantly improves the practicality and economic benefits of the model. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a multi-modal data serialization schematic diagram of the application;

[0059] Figure 2 is a time sequence model structure schematic diagram of the application;

[0060] Figure 3 is a model training flow schematic diagram of the application; DETAILED DESCRIPTION

[0061] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the specification herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the description and claims of this application as well as the above discussion of the related art are intended to be illustrative only and not restrictive. The word "comprising" and variations such as "comprise" or "comprises" as used in the specification are to be interpreted in an inclusive sense, that is, as "including, but not limited to." The words "first," "second," and "third," and the like, as used in the specification, are to be interpreted as distinguishing between objects, and not to require a specific order or sequence of inclusion.

[0062] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase "in an embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. It is expressly understood that any of the embodiments described herein can be incorporated into any other embodiment.

[0063] The embodiment of the application provides a car risk control time sequence prediction method based on a large model technology, as shown in Figures 1-3 Fig. 1, which comprises S1: fusion and serialization processing of multi-source heterogeneous data to obtain time sequence data; the multi-source heterogeneous data comprises credit investigation video data, vehicle GPS data, periodic return visit audio data and repayment behavior data;

[0064] S2: constructing a time series prediction model based on an attention mechanism, receiving past P time point input contexts of time series data, and outputting overdue prediction results of future F time points;

[0065] S3: optimizing a sampling strategy of the time series prediction model through a reinforcement learning method to improve model prediction performance.

[0066] It should be noted that, since existing risk control methods rely on expert experience and blacklists, there are problems of timeliness, privacy, and inability to adapt to dynamic changes in risk patterns. Although methods based on machine learning and deep learning can improve risk identification accuracy, they rely on a large amount of labeled data, have a lag, and have difficulty effectively fusing and extracting key information when processing multi-source heterogeneous data, limiting the accuracy and robustness of the prediction model. The present scheme constructs a comprehensive, accurate, and efficient time series prediction framework through multi-source heterogeneous data fusion, fine time series preprocessing, attention mechanism modeling, and reinforcement learning dynamic optimization of the sampling strategy, not only significantly improving the accuracy, timeliness, and robustness of the prediction, reducing data redundancy, and improving data processing and model training efficiency, but also enhancing the universality and adaptability of the model in different scenarios, while reducing the dependence on historical data, enabling timely risk pattern mining and customer risk behavior prediction, providing reliable, intelligent, and more timely and cost-effective decision support for the automotive risk control field.

[0067] Specifically, in the present embodiment, the present scheme mainly uses a large model technology-based automotive risk control time series prediction method, fuses multi-source heterogeneous data, including credit review video, vehicle GPS, regular visit audio, and repayment behavior data, and serializes the processing to construct a comprehensive time series data foundation. Then, based on the attention mechanism, a prediction model is constructed to accurately capture the complex relationships between data and output future overdue prediction results. Finally, through the reinforcement learning method, the sampling strategy is optimized to significantly improve the model prediction performance.

[0068] As shown in the further preferred embodiment of the present application, Figures 1-3 the step S1 includes the following sub-steps:

[0069] S11: adaptively sampling and fixedly sampling the credit review video data to obtain candidate video frames;

[0070] S12: extracting stay points and performing multi-modal conversion on the vehicle GPS data to obtain GPS text data and GPS image data;

[0071] S13: performing feature extraction on the regular visit audio data to obtain audio feature data;

[0072] S14: performing standardized text conversion on the repayment behavior data to generate repayment record text;

[0073] S15: Modal alignment and time alignment of image, text, and audio data are performed to form a unified time series input.

[0074] In this embodiment, adaptive and fixed sampling are performed on the credit investigation video data, stay point extraction and multi-modal conversion are performed on the vehicle GPS data, features are extracted from the regular follow-up audio data, standardized text conversion is performed on the repayment behavior data, and finally modal alignment and time alignment of image, text, and audio data are realized to form a unified and comprehensive time series input. This process not only integrates key information from multiple heterogeneous data sources, but also ensures the consistency of data in time and modal, providing a high-quality data basis for subsequent time series prediction models, thereby helping to improve the accuracy and reliability of the prediction model.

[0075] In further preferred embodiments of the application, as shown in Figures 1-3 Step S11 includes:

[0076] The overlap rate of adjacent video frames is calculated, and given two video frames F i and F j , the calculation formula of the overlap rate is as follows:

[0077] In this embodiment, by calculating the overlap rate of adjacent video frames, representative and information-rich candidate video frames can be effectively selected. This operation not only reduces data redundancy and improves data processing efficiency, but also ensures that the selected video frames accurately reflect the key information in the credit investigation process, providing strong support for building more accurate automobile risk control time series prediction models.

[0078] In further preferred embodiments of the application, as shown in Figures 1-3 Step S12 includes:

[0079] Stay point extraction: based on time threshold and distance threshold, all points that meet the given threshold conditions in the n-day GPS points are selected as the candidate stay point set of the user;

[0080] Stay point fusion: the mean fusion of continuous GPS points in the same stay point area is performed, that is, from the first stay point based on time series sorting, the mean fusion of stay points within the same distance threshold range is performed as a new stay point, and finally the stay point set of the user in the past n days is obtained;

[0081] Multi-modal conversion: the n-day stay point set is concatenated into GPS text data and drawn as a GPS heat map image to obtain GPS image data.

[0082] In this embodiment, through the continuous stop point extraction, fusion and multi-modal conversion operation, the key information in the vehicle GPS data is effectively refined, first, the candidate stop points are screened based on the time and distance threshold, the accuracy of the data is ensured, then, the mean fusion is performed on the continuous GPS points in the same area, the data is further simplified and the core information is retained, finally, the stop point data is converted into text and image form through multi-modal conversion, the data expression is enriched, a more comprehensive and intuitive input is provided for the subsequent time series prediction model, and the prediction performance and accuracy of the model are improved.

[0083] As shown in the further preferred embodiment of the present application, Figures 1-3 Step S13 comprises:

[0084] The audio data is preprocessed by noise reduction and framing, the mel frequency cepstral coefficient and short-time energy spectrum are calculated, and the audio feature tensor is obtained.

[0085] In this embodiment, through the continuous noise reduction, framing preprocessing, and calculation of the mel frequency cepstral coefficient and short-time energy spectrum, the key features in the regular follow-up audio data are effectively extracted, and a structured audio feature tensor is formed. This process not only removes noise interference, but also captures the frequency and energy change characteristics of the audio signal, providing high-quality and information-rich audio input for the subsequent time series prediction model, which helps to improve the understanding and prediction ability of the model for audio data.

[0086] As shown in the further preferred embodiment of the present application, Figures 1-3 Step S14 comprises:

[0087] The repayment record is converted into structured text with the format "timestamp: repayment amount→repayment amount", and the repayment behavior text data is generated.

[0088] In this embodiment, by converting the repayment record into structured text of a specific format, the repayment behavior data is effectively organized and standardized, making the repayment data clearer and easier to process, and providing a unified and standardized text input for the subsequent time series prediction model, which helps to improve the understanding and analysis accuracy of the model for the repayment behavior data, and further enhances the accuracy and reliability of the prediction results.

[0089] As shown in the further preferred embodiment of the present application, Figures 1-3 Step S15 comprises:

[0090] Time alignment: taking the second-level timestamp as the reference, mapping the video frame, GPS point, audio segment and repayment record to the unified time axis, and filling in the missing values through the context prediction of the large model;

[0091] Chunked serialization: The multi-modal data is chunked by fixed time windows, and each chunk is spliced in time sequence to form a multi-modal sequence, and position encoding is added to retain the time sequence information.

[0092] In this embodiment, through the consecutive time alignment and chunked serialization operations, the precise mapping and integration of multi-source heterogeneous data such as video frames, GPS points, audio segments, and repayment records on a unified time axis are realized. Time alignment ensures the consistency of each data source in the time dimension, and missing values are completed through large model context prediction, improving the completeness and accuracy of the data. Chunked serialization further chunks the multi-modal data by fixed time windows, splices them in time sequence within the chunk, and adds position encoding to retain the time sequence information, providing a clear structure and rich information input for subsequent time series prediction models, which helps to improve the prediction performance and accuracy of the model.

[0093] In further preferred embodiments of the application, as shown in Figures 1-3 The time series prediction model of step S2 is based on a large model architecture, specifically including:

[0094] A time series prediction model is constructed, which takes the data of the past P time points in the time series as input context information to predict the potential delinquency status of the customer at the future F time points:

[0095]

[0096] The prediction accuracy is evaluated by the metric corresponding to the current delinquency day index of the repayment behavior data, which quantifies the closeness between the predicted value and the true value, and the mean absolute error is used as this metric:

[0097]

[0098] Step 1. In the time series preprocessing link, to generate the input token for the transformation layer, first divide the input time series data into continuous and non-overlapping segments, and after processing by the residual block, convert it into a vector adapted to the model dimension. At the same time, introduce a binary format fill mask, where 1 represents the corresponding input value is masked, and 0 represents it is not masked. After the above processing, the input data is divided into segments of length n, and the i-th segment is denoted as The corresponding mask segmentation is denoted as The i-th input token is denoted as:

[0099]

[0100] where PE i represents the i-th position encoding, and represents element-wise multiplication, and there are a total of an input token;

[0101] Step 2. Design the stacked transformer layer as the core structure of the model, using the standard causal multi-head self-attention mechanism, ensuring that each output token can only focus on the input tokens arranged before it when performing attention calculation, which is represented by the following formula:

[0102]

[0103] where ST represents the stacked transformer layer, t i represents the token at position i, the minimum mask at the i-th position;

[0104] Step 3. Design the output layer to map the output token to the predicted value, the model is trained in decoder mode, ensuring that each output token predicts the time series part that follows the last input segment corresponding to itself; set the output segment length to h, use another residual block to map the output token to the predicted value, which is represented by the following formula:

[0105]

[0106] As above, all data encoding in y 1:ni is completed, and it is used to predict the sequence y ni+1:ni+h at the next h time points;

[0107] This model aims to predict the overdue days in the repayment behavior data, and uses mean square error as the loss function to measure the difference between the predicted value and the true value and optimize the model during training:

[0108]

[0109] W is the number of samples in a batch, is the predicted value of the next segment of the time series, y ni+1:ni+h is the corresponding actual observation value.

[0110] In this embodiment, the model takes the data of the past P time points in the time series as input, accurately capturing historical information, which is used to predict the potential overdue status of the customer at the future F time points, effectively predicting the future trend;

[0111] Secondly, by taking the mean absolute error as the measurement standard of prediction accuracy, the model can quantify the difference between the predicted value and the true value, providing a clear direction for optimization; in terms of model architecture, the preprocessing link effectively generates input tokens that are adapted to the dimensions of the model through segmentation, residual block processing and the introduction of padding masks, laying a foundation for subsequent calculations;

[0112] The core structure adopts stacked transformer layers combined with causal multi-head self-attention mechanism, ensuring the rationality of the causal relationship and attention allocation of the model when processing time series data.

[0113] The design of the output layer further maps the output token to the predicted value, adopts the decoder mode for training, and ensures the continuity and accuracy of the prediction. Finally, the model takes the overdue days in the predicted repayment behavior data as the target, adopts the mean square error as the loss function for optimization, effectively improves the prediction accuracy, and provides an efficient and accurate overdue prediction tool for the automobile risk control field.

[0114] As shown in the further preferred embodiment of the application, Figures 1-3 Step S3 includes the following sub-steps:

[0115] S31: define a set of time series sampling strategies, including weighted random sampling: let w(t) = mu (T-t) The weight of the data point at time t, wherein w(t) = mu (T-t) Close to but less than 1, T is the selected prediction time point, and the sampling probability is defined as

[0116] S32: build a reinforcement learning framework: the state space S is the prediction accuracy of the model within the time window; the action space A is the set of selectable sampling strategies;

[0117] The reward function is:

[0118] Wherein A t Indicates the accuracy of the model in the time window [t, t+Delta t], A threshold Indicates the preset accuracy threshold, and alpha indicates the hyperparameter controlling the slope of the reward function.

[0119] The strategy function is:

[0120] Wherein a t Indicates the action selected at the time step, s t Indicates the state at time step t, i.e. the accuracy of the sample prediction in the time window [t, t+Delta t];

[0121] S33: update the strategy through a deep Q network, and the algorithm formula is as follows:

[0122]

[0123] S34: dynamically select the sampling strategy according to the optimized strategy, retrain the time series prediction model until the model accuracy converges.

[0124] In this embodiment, the sampling strategy can be dynamically adjusted according to the prediction accuracy of the model in different time windows, so as to more effectively utilize data resources and improve the prediction performance of the model. Through the combination of the reinforcement learning framework and the deep Q network, the model can autonomously learn the optimal sampling strategy, and then continuously optimize in the retraining process until a high prediction accuracy is reached. This not only improves the adaptability and robustness of the model, but also provides a more accurate and efficient solution for time series prediction in the field of automobile risk control.

[0125] It should be noted that, for the foregoing embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0126] In several embodiments provided in the present application, it should be understood that the disclosed device can be implemented by other ways. For example, the device embodiments described above are only schematic, for example, the division of the above units can have another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, which can be electrical or other form.

[0127] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment scheme according to actual needs.

[0128] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the protection scope of the application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art can still combine, delete or make other adjustments to the features of the embodiments of the present application according to the circumstances without making creative labor, so as to obtain different other technical solutions which do not deviate from the concept of the present application in essence. These technical solutions also belong to the scope of the present application.

Claims

1. A method for predicting automobile risk control time series based on large model technology, characterized in that: include: S1: Fusion and serialization of multi-source heterogeneous data to obtain time series data; multi-source heterogeneous data includes credit review video data, vehicle GPS data, regular return visit audio data, and repayment behavior data; S2: Build a time series prediction model based on the attention mechanism, receive the input context of the past P time points of the time series data, and output the overdue prediction results for the next F time points; S3: Optimize the sampling strategy of the time series prediction model through reinforcement learning methods to improve the model prediction performance.

2. The automobile risk control time series prediction method based on large model technology according to claim 1 is characterized in that: Step S1 includes the following sub-steps: S11: Perform adaptive sampling and fixed sampling on the credit-reviewed video data to obtain candidate video frames; S12: performing stop point extraction and multimodal conversion on the vehicle GPS data to obtain GPS text data and GPS image data; S13: extracting features from the periodic return visit audio data to obtain audio feature data; S14: performing standardized text conversion on the repayment behavior data to generate a repayment record text; S15: Perform modality alignment and time alignment on image, text, and audio data to form a unified time series input.

3. The automobile risk control time series prediction method based on large model technology according to claim 2 is characterized in that: Step S11 includes: Calculate the overlap ratio of adjacent video frames. Given two video frames F i and F j , its overlap ratio The calculation formula is as follows:

4. The automobile risk control time series prediction method based on large model technology according to claim 2 is characterized in that: Step S12 includes: Stay point extraction: Based on time threshold and distance threshold, all points that meet the given threshold conditions in the GPS points in the past n days are selected as the user's candidate stay point set; Stay point fusion: This method performs mean fusion on consecutive GPS points in the same stay point area. Starting from the first stay point sorted by time series, all stay points within the same distance threshold are mean fused as new stay points, ultimately obtaining a set of the user's stay points over the past n days. Multimodal conversion: The set of n-day stay points is spliced ​​into GPS text data and plotted as a GPS thermal image to obtain GPS image data.

5. The automobile risk control time series prediction method based on large model technology according to claim 2 is characterized in that: Step S13 includes: The audio data is subjected to denoising and frame preprocessing, and the Mel-frequency cepstral coefficients and short-time energy spectrogram are calculated to obtain the audio feature tensor.

6. The automobile risk control time series prediction method based on large model technology according to claim 2 is characterized in that: Step S14 includes: Convert repayment records into structured text in the format of "timestamp: due amount → actual amount repaid" to generate repayment behavior text data.

7. The automobile risk control time series prediction method based on large model technology according to claim 2 is characterized in that: Step S15 includes: Time alignment: Using second-level timestamps as a benchmark, we map video frames, GPS points, audio clips, and repayment records to a unified timeline. Missing values ​​are complemented through contextual predictions from a large model. Block serialization: Divide multimodal data into blocks according to fixed time windows, splice each block into a multimodal sequence in chronological order, and add position encoding to preserve the time order information.

8. The automobile risk control time series prediction method based on large model technology according to claim 1 is characterized in that: The time series prediction model described in step S2 is based on a large model architecture and specifically includes: Build a time series prediction model. This time series prediction model uses the data of the past P time points in the time series as input context information to predict the potential overdue status of customers at the next F time points: The prediction accuracy is evaluated by the metric corresponding to the current overdue days indicator of the repayment behavior data. This metric is used to quantify the closeness between the predicted value and the true value. The mean absolute error is used as this metric: Step 1. In the time series preprocessing phase, in order to generate the input tokens required by the transformation layer, the input time series data is first divided into continuous and non-overlapping slices. Each slice is processed by the residual block and converted into a vector adapted to the model dimension. At the same time, a binary format padding mask is introduced, where 1 indicates that the corresponding input value is masked and 0 indicates that it is not masked. After the above processing, the input data is decomposed into slices of length n, and the i-th slice is recorded as The corresponding mask segmentation is recorded as The i-th input token is represented as: Among them, PE i represents the i-th position code, ⊙ represents element-by-element multiplication, totaling Input token; Step 2. Design a stacked transformer layer as the core structure of the model and use a standard causal multi-head self-attention mechanism to ensure that each output token can only focus on the input token that comes before it in the sequence when performing attention calculation. The formula is as follows: Among them, ST represents the stacked transformation layer, t i Represents the token at position i, represents the minimum mask at the i-th position; Step 3. Design the output layer to map the output tokens to the predicted values. The model is trained in decoder mode to ensure that each output token prediction is the time series part immediately following the last input slice corresponding to it. Assume that the output slice length is h, and use another residual block to map the output token to the predicted value. The formula is as follows: Complete all y as above 1:ni Encode the data in and use it to predict the sequence y of the next h time points ni+1:ni+h ; This model aims to predict the number of overdue days from repayment behavior data. During training, the mean square error is used as the loss function to measure the difference between the predicted value and the true value and optimize the model: W is the number of samples in a batch, is the model’s predicted value for the next slice of the time series, y ni+1:ni+h is the corresponding actual observed value.

9. The automobile risk control time series prediction method based on large model technology according to claim 1 is characterized in that: Step S3 includes the following sub-steps: S31: Define a set of time series sampling strategies, including weighted random sampling: Let w(t) = μ (T-t) is the weight of the data point at time t, where μ is close to but less than 1, T is the selected prediction time point, and the sampling probability is defined as S32: Construct a reinforcement learning framework: the state space S is the prediction accuracy of the model within the time window; the action space A is the set of optional sampling strategies; The reward function is Among them A t A represents the accuracy of the model in the time window [t, t+Δt], threshold represents the preset accuracy threshold, and α represents the hyperparameter that controls the slope of the reward function; The policy function is where a t represents the action selected at time step, s t Represents the state at time step t, that is, the accuracy of sample prediction in the time window [t, t+Δt]; S33: Update the strategy through the deep Q network. The algorithm formula is as follows: S34: Dynamically select the sampling strategy according to the optimization strategy and retrain the time series prediction model until the model accuracy converges.

Images